Red Hat reports a 70% cut in frontier model costs

Red Hat, a maker of open software, showed OpenShift 5 and open tools for AI agents before KubeCon North America. Steve Watt, a Red Hat vice president, says a three to four week pilot cut frontier model costs by 70%. Most developer traffic moved to GLM, an open model with a very low cost for Red Hat.

How it works

EnMaaS is an open stack to control the cost of AI models. Each programmer gets a key that shows costs and sets a quota of hundreds of dollars each month. It has these parts:

• Praxis: a gateway that checks prompts and sends them to the correct model.
• llm-d: software that operates open models on many GPU clusters.
• Openshell: an Nvidia sandbox that limits the tasks of an agent.

The numbers

Steve Watt gave one result: the pilot cut frontier model costs by 70%. Frontier models are top closed models. Red Hat also says that its Lightwell project made patches for more than 400 unknown vulnerabilities in Java libraries.

What is not known

Red Hat gave no data on the quality of GLM output or on the pilot. Red Hat is the only source for the 400 vulnerabilities. It gave no test results for OpenShift 5.

Insight: Rival firms build sandboxes for agents

Red Hat puts the Nvidia sandbox Openshell in its stack. AWS released Strands Box in October 2026, and Alibaba gave OpenSandbox to the Agentic AI Foundation. Each seems to want to control the layer that all agents use. The first sandbox that teams use can set the rules. Before you select one, make sure that your agents can move to a different one.